This is an embarrassment actually, after the excellent analysis in the 
PR trail.

We've (I've) been using the wrong sign convention for the nu parameter 
in comp_ellint_3 and in ellint_3 in tr1 and now std for years.  In my 
defence the World seems about evenly split on this with GSL, 
Abramowitz&Stegun (and DLMF) and Carlson (that's what actually got me 
here) on the +nu side, and tr1/std, boost on the other -nu side.

We need to get this into 8.

I'm not sure about backporting to 7.  I don't think we should change 
this breaking in the middle of a release series.

I'll write a Mea Culpa warning about changing ellint_3, comp_ellint_3 in 
the release notes and in the libstdc++ pages and the tr29124 conformance 
page later.

This builds and passes on x86-64-linux.

OK for 8?

Ed